Agents and MCP

What the MCP server lets an agent do.

What the server lets an agent do

KonfAI MCP exposes KonfAI to an MCP client (Claude Code, Codex, Studio) as fifty tools, one per object: inspect a dataset, design and write a config, check it without launching anything (up to one real forward and backward pass), launch and follow jobs, compare runs, use a published app, fine-tune it, package a session as an app, inspect a result with Scout, read FidOpt's catalogue. Every run is a job; most answers name the valid next calls.

pip install "konfai[imaging]" konfai-apps konfai-mcp
konfai-mcp            # stdio transport by default

[mcp_servers.konfai]
command = "/path/to/venv/bin/konfai-mcp"
tool_timeout_sec = 3600

[mcp_servers.konfai.env]
KONFAI_MCP_WORKSPACES_ROOT = "/path/to/workspaces"

The tools by stage

Stage	Tools
Orient	`describe_konfai_capabilities`, `describe_config_schema`, `describe_extension_points`
Dataset	`inspect_dataset` (and `browse=True`), `read_file`, `preview_volume`, `prepare_dataset_aliases`
Author	`design_config_strategy`, `initialize_session`, `write_workflow_config`, `write_session_file`
Validate	`check_config` (`review`, `instantiate`, `setup`, `train_step`)
Run and monitor	`run_train`, `run_prediction`, `run_evaluation`, `run_transform` (`dry_run=True` first), `get_job_status`, `read_live_metrics`, `leaderboard`
Use an app	`list_apps`, `describe_app`, `run_app` (`infer`, `evaluate`, `uncertainty`, `pipeline`), `fetch_app`
Adapt and package	`fine_tune_app`, `run_resume`, `package_app_from_session`, `app_sources`

The six entry tools a client loads first: describe_konfai_capabilities, list_apps, run_app, initialize_session, write_workflow_config, summarize_session. The complete table, one sentence per tool: MCP on this site; the server guide: MCP workflows.